Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore how YouTube is revolutionizing its recommendation system by adapting Gemini, Google's large language model, to serve billions of daily active users in this 23-minute conference talk. Learn about the fundamental shift from traditional large embedding models to foundation models for powering video recommendations at massive scale. Discover the engineering challenges and solutions involved in teaching general-purpose LLMs to understand YouTube's unique video ecosystem, including the development of SemanticID - a specialized "language" for YouTube videos that enables better generalization in recommendation ranking. Examine the process of adapting Gemini checkpoints to comprehend SemanticID and implement generative video retrieval using prompts. Gain insights into the broader LLM-led transformation of recommendation systems and feeds across large consumer applications, understanding how this technology shift extends beyond search to fundamentally change how billions of users discover content. The presentation covers practical implementation details of building a recommendation engine on top of Gemini, drawing from real-world experience in scaling generative recommendation systems from research concepts to consumer-facing products serving YouTube's massive user base.
